In [1]:
cd ../..
/code
In [2]:
%run "source/config/notebook_settings.py"
import os
import mlflow
from mlflow.tracking import MlflowClient
from helpsk.utility import read_pickle
import helpsk as hlp

from source.library.utilities import Timer, log_info, get_config

config = get_config()
mlflow_uri = config['MLFLOW']['URI']
log_info(f"MLFlow URI: {mlflow_uri}")

client = MlflowClient(tracking_uri=mlflow_uri)
2022-06-20 02:29:10 - INFO     | MLFlow URI: http://mlflow_server:1235

Get Latest Experiment Run from MLFlow¶

In [3]:
# Get the production model version and actual model
production_model_info = client.get_latest_versions(name=config['MLFLOW']['MODEL_NAME'], stages=['Production'])
assert len(production_model_info) == 1
production_model_info = production_model_info[0]
production_model = read_pickle(client.download_artifacts(
    run_id=production_model_info.run_id,
    path='model/model.pkl'
))
log_info(f"Production Model Version: {production_model_info.version}")
2022-06-20 02:29:10 - INFO     | Production Model Version: 2
In [4]:
# get experiment and latest run info
credit_experiment = client.get_experiment_by_name(name=config['MLFLOW']['EXPERIMENT_NAME'])
runs = client.list_run_infos(experiment_id=credit_experiment.experiment_id)
latest_run = runs[np.argmax([x.start_time for x in runs])]
In [5]:
yaml_path = client.download_artifacts(run_id=latest_run.run_id, path='experiment.yaml')
results = hlp.sklearn_eval.MLExperimentResults.from_yaml_file(yaml_file_name = yaml_path)
In [6]:
# get the best estimator from the BayesSearchCV
best_estimator = read_pickle(client.download_artifacts(
    run_id=latest_run.run_id,
    path='model/model.pkl'
))
In [7]:
best_estimator.model
Out[7]:
Pipeline(steps=[('prep',
                 ColumnTransformer(transformers=[('numeric',
                                                  Pipeline(steps=[('imputer',
                                                                   TransformerChooser(transformer=SimpleImputer(strategy='median'))),
                                                                  ('scaler',
                                                                   TransformerChooser()),
                                                                  ('pca',
                                                                   TransformerChooser(transformer=PCA(n_components='mle')))]),
                                                  ['duration', 'credit_amount',
                                                   'installment_commitment',
                                                   'residence_since', 'age',
                                                   'existing_credi...
                                                   'personal_status',
                                                   'other_parties',
                                                   'property_magnitude',
                                                   'other_payment_plans',
                                                   'housing', 'job',
                                                   'own_telephone',
                                                   'foreign_worker'])])),
                ('model',
                 ExtraTreesClassifier(bootstrap=True, criterion='entropy',
                                      max_depth=99,
                                      max_features=0.031837350792579364,
                                      max_samples=0.9248344222191298,
                                      min_samples_leaf=4, min_samples_split=16,
                                      n_estimators=1235, random_state=42))])
In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.
Pipeline(steps=[('prep',
                 ColumnTransformer(transformers=[('numeric',
                                                  Pipeline(steps=[('imputer',
                                                                   TransformerChooser(transformer=SimpleImputer(strategy='median'))),
                                                                  ('scaler',
                                                                   TransformerChooser()),
                                                                  ('pca',
                                                                   TransformerChooser(transformer=PCA(n_components='mle')))]),
                                                  ['duration', 'credit_amount',
                                                   'installment_commitment',
                                                   'residence_since', 'age',
                                                   'existing_credi...
                                                   'personal_status',
                                                   'other_parties',
                                                   'property_magnitude',
                                                   'other_payment_plans',
                                                   'housing', 'job',
                                                   'own_telephone',
                                                   'foreign_worker'])])),
                ('model',
                 ExtraTreesClassifier(bootstrap=True, criterion='entropy',
                                      max_depth=99,
                                      max_features=0.031837350792579364,
                                      max_samples=0.9248344222191298,
                                      min_samples_leaf=4, min_samples_split=16,
                                      n_estimators=1235, random_state=42))])
ColumnTransformer(transformers=[('numeric',
                                 Pipeline(steps=[('imputer',
                                                  TransformerChooser(transformer=SimpleImputer(strategy='median'))),
                                                 ('scaler',
                                                  TransformerChooser()),
                                                 ('pca',
                                                  TransformerChooser(transformer=PCA(n_components='mle')))]),
                                 ['duration', 'credit_amount',
                                  'installment_commitment', 'residence_since',
                                  'age', 'existing_credits',
                                  'num_dependents']),
                                ('non_numeric',
                                 Pipeline(steps=[('encoder',
                                                  TransformerChooser(transformer=CustomOrdinalEncoder()))]),
                                 ['checking_status', 'credit_history',
                                  'purpose', 'savings_status', 'employment',
                                  'personal_status', 'other_parties',
                                  'property_magnitude', 'other_payment_plans',
                                  'housing', 'job', 'own_telephone',
                                  'foreign_worker'])])
['duration', 'credit_amount', 'installment_commitment', 'residence_since', 'age', 'existing_credits', 'num_dependents']
TransformerChooser(transformer=SimpleImputer(strategy='median'))
SimpleImputer(strategy='median')
SimpleImputer(strategy='median')
TransformerChooser()
TransformerChooser(transformer=PCA(n_components='mle'))
PCA(n_components='mle')
PCA(n_components='mle')
['checking_status', 'credit_history', 'purpose', 'savings_status', 'employment', 'personal_status', 'other_parties', 'property_magnitude', 'other_payment_plans', 'housing', 'job', 'own_telephone', 'foreign_worker']
TransformerChooser(transformer=CustomOrdinalEncoder())
CustomOrdinalEncoder()
CustomOrdinalEncoder()
ExtraTreesClassifier(bootstrap=True, criterion='entropy', max_depth=99,
                     max_features=0.031837350792579364,
                     max_samples=0.9248344222191298, min_samples_leaf=4,
                     min_samples_split=16, n_estimators=1235, random_state=42)

Training & Test Data Info¶

In [8]:
client.download_artifacts(run_id=latest_run.run_id, path='x_train.pkl')
Out[8]:
'/code/mlflow-artifact-root/1/7bdbdabe0cfb4219a3ca8b682166db0a/artifacts/x_train.pkl'
In [9]:
with Timer("Loading training/test datasets"):
    X_train = pd.pandas.read_pickle(client.download_artifacts(run_id=latest_run.run_id, path='x_train.pkl'))
    X_test = pd.pandas.read_pickle(client.download_artifacts(run_id=latest_run.run_id, path='x_test.pkl'))
    y_train = pd.pandas.read_pickle(client.download_artifacts(run_id=latest_run.run_id, path='y_train.pkl'))
    y_test = pd.pandas.read_pickle(client.download_artifacts(run_id=latest_run.run_id, path='y_test.pkl'))
2022-06-20 02:29:11 - INFO     | Timer Started: Loading training/test datasets
2022-06-20 02:29:11 - INFO     | Timer Finished: (0.02 seconds)
In [10]:
log_info(X_train.shape)
log_info(len(y_train))

log_info(X_test.shape)
log_info(len(y_test))
2022-06-20 02:29:11 - INFO     | (800, 20)
2022-06-20 02:29:11 - INFO     | 800
2022-06-20 02:29:11 - INFO     | (200, 20)
2022-06-20 02:29:11 - INFO     | 200
In [11]:
np.unique(y_train, return_counts=True)
Out[11]:
(array([0, 1]), array([559, 241]))
In [12]:
np.unique(y_train, return_counts=True)[1] / np.sum(np.unique(y_train, return_counts=True)[1])
Out[12]:
array([0.69875, 0.30125])
In [13]:
np.unique(y_test, return_counts=True)[1] / np.sum(np.unique(y_test, return_counts=True)[1])
Out[13]:
array([0.705, 0.295])

Cross Validation Results¶

Best Scores/Params¶

In [14]:
log_info(f"Best Score: {results.best_score}")
2022-06-20 02:29:11 - INFO     | Best Score: 0.7700760706900486
In [15]:
log_info(f"Best Params: {results.best_params}")
2022-06-20 02:29:11 - INFO     | Best Params: {'model': 'ExtraTreesClassifier()', 'max_features': 0.031837350792579364, 'max_depth': 99, 'n_estimators': 1235, 'min_samples_split': 16, 'min_samples_leaf': 4, 'max_samples': 0.9248344222191298, 'criterion': 'entropy', 'imputer': "SimpleImputer(strategy='median')", 'scaler': 'None', 'pca': "PCA('mle')", 'encoder': 'CustomOrdinalEncoder()'}
In [16]:
# Best model from each model-type.
df = results.to_formatted_dataframe(return_style=False, include_rank=True)
df["model_rank"] = df.groupby("model")["roc_auc Mean"].rank(method="first", ascending=False)
df.query('model_rank == 1')
Out[16]:
rank roc_auc Mean roc_auc 95CI.LO roc_auc 95CI.HI model C max_features max_depth n_estimators min_samples_split min_samples_leaf max_samples criterion learning_rate min_child_weight subsample colsample_bytree colsample_bylevel reg_alpha reg_lambda num_leaves imputer scaler pca encoder model_rank
9 1 0.77 0.73 0.81 ExtraTreesClassifier() NaN 0.03 99.00 1235.00 16.00 4.00 0.92 entropy NaN NaN NaN NaN NaN NaN NaN NaN SimpleImputer(strategy='median') None PCA('mle') CustomOrdinalEncoder() 1.00
3 2 0.76 0.71 0.81 LogisticRegression() 0.01 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN SimpleImputer() StandardScaler() PCA('mle') OneHotEncoder() 1.00
10 3 0.76 0.70 0.82 RandomForestClassifier() NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN SimpleImputer() None None OneHotEncoder() 1.00
16 7 0.75 0.70 0.79 XGBClassifier() NaN NaN 5.00 1157.00 NaN NaN NaN NaN 0.02 3.00 0.69 0.50 0.73 0.03 2.91 NaN SimpleImputer(strategy='median') None PCA('mle') CustomOrdinalEncoder() 1.00
21 16 0.73 0.68 0.78 LGBMClassifier() NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN 0.61 0.98 NaN 8.89 7.31 243.00 SimpleImputer(strategy='most_frequent') None PCA('mle') OneHotEncoder() 1.00
In [17]:
results.to_formatted_dataframe(return_style=True,
                               include_rank=True,
                               num_rows=500)
Out[17]:
rank roc_auc Mean roc_auc 95CI.LO roc_auc 95CI.HI model C max_features max_depth n_estimators min_samples_split min_samples_leaf max_samples criterion learning_rate min_child_weight subsample colsample_bytree colsample_bylevel reg_alpha reg_lambda num_leaves imputer scaler pca encoder
1 0.770 0.732 0.808 ExtraTreesClassifier() <NA> 0.032 99.000 1,235.000 16.000 4.000 0.925 entropy <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer(strategy='median') None PCA('mle') CustomOrdinalEncoder()
2 0.762 0.714 0.810 LogisticRegression() 0.014 <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer() StandardScaler() PCA('mle') OneHotEncoder()
3 0.761 0.702 0.820 RandomForestClassifier() <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer() None None OneHotEncoder()
4 0.759 0.712 0.806 LogisticRegression() 0.000 <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer(strategy='most_frequent') MinMaxScaler() None OneHotEncoder()
5 0.757 0.711 0.803 LogisticRegression() <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer() StandardScaler() None OneHotEncoder()
6 0.749 0.724 0.774 ExtraTreesClassifier() <NA> 0.563 71.000 1,725.000 49.000 16.000 0.956 gini <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer(strategy='median') None None CustomOrdinalEncoder()
7 0.747 0.704 0.789 XGBClassifier() <NA> <NA> 5.000 1,157.000 <NA> <NA> <NA> <NA> 0.018 3.000 0.694 0.501 0.726 0.033 2.910 <NA> SimpleImputer(strategy='median') None PCA('mle') CustomOrdinalEncoder()
8 0.747 0.711 0.782 XGBClassifier() <NA> <NA> 4.000 1,414.000 <NA> <NA> <NA> <NA> 0.013 1.000 0.841 0.624 0.542 0.784 1.243 <NA> SimpleImputer() None PCA('mle') CustomOrdinalEncoder()
9 0.746 0.698 0.794 ExtraTreesClassifier() <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer() None None OneHotEncoder()
10 0.745 0.703 0.787 RandomForestClassifier() <NA> 0.757 44.000 745.000 33.000 6.000 0.608 gini <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer() None None CustomOrdinalEncoder()
11 0.742 0.691 0.793 XGBClassifier() <NA> <NA> 2.000 1,671.000 <NA> <NA> <NA> <NA> 0.021 2.000 0.657 0.591 0.780 0.026 3.081 <NA> SimpleImputer(strategy='median') None None OneHotEncoder()
12 0.738 0.727 0.748 ExtraTreesClassifier() <NA> 0.861 52.000 1,995.000 33.000 19.000 0.651 gini <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer(strategy='most_frequent') None None CustomOrdinalEncoder()
13 0.737 0.716 0.757 RandomForestClassifier() <NA> 0.528 70.000 1,003.000 37.000 19.000 0.530 gini <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer(strategy='median') None PCA('mle') CustomOrdinalEncoder()
14 0.734 0.709 0.758 RandomForestClassifier() <NA> 0.422 68.000 1,348.000 4.000 39.000 0.859 entropy <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer(strategy='median') None PCA('mle') CustomOrdinalEncoder()
15 0.733 0.705 0.760 ExtraTreesClassifier() <NA> 0.666 19.000 1,346.000 6.000 44.000 0.852 gini <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer(strategy='median') None PCA('mle') OneHotEncoder()
16 0.732 0.682 0.782 LGBMClassifier() <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> 0.612 0.979 <NA> 8.891 7.306 243.000 SimpleImputer(strategy='most_frequent') None PCA('mle') OneHotEncoder()
17 0.731 0.689 0.772 LGBMClassifier() <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> 0.338 0.617 <NA> 10.572 26.393 255.000 SimpleImputer(strategy='median') None PCA('mle') OneHotEncoder()
18 0.729 0.645 0.814 LogisticRegression() 0.000 <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer(strategy='most_frequent') StandardScaler() PCA('mle') OneHotEncoder()
19 0.727 0.694 0.759 RandomForestClassifier() <NA> 0.940 40.000 899.000 3.000 16.000 0.696 entropy <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer(strategy='median') None None CustomOrdinalEncoder()
20 0.724 0.685 0.762 XGBClassifier() <NA> <NA> 13.000 1,222.000 <NA> <NA> <NA> <NA> 0.080 5.000 0.825 0.700 0.977 0.020 1.105 <NA> SimpleImputer(strategy='median') None PCA('mle') CustomOrdinalEncoder()
21 0.723 0.705 0.741 LogisticRegression() 0.000 <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer(strategy='median') StandardScaler() None CustomOrdinalEncoder()
22 0.719 0.628 0.810 LGBMClassifier() <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer() None None OneHotEncoder()
23 0.718 0.682 0.753 LGBMClassifier() <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> 0.925 0.859 <NA> 14.754 33.903 400.000 SimpleImputer(strategy='median') None PCA('mle') CustomOrdinalEncoder()
24 0.714 0.648 0.779 XGBClassifier() <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer() None None OneHotEncoder()
25 0.702 0.683 0.721 LGBMClassifier() <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> 0.374 0.990 <NA> 16.223 18.789 378.000 SimpleImputer() None PCA('mle') CustomOrdinalEncoder()
In [18]:
results.to_formatted_dataframe(query='model == "RandomForestClassifier()"', include_rank=True)
Out[18]:
rank roc_auc Mean roc_auc 95CI.LO roc_auc 95CI.HI max_features max_depth n_estimators min_samples_split min_samples_leaf max_samples criterion imputer pca encoder
1 0.761 0.702 0.820 <NA> <NA> <NA> <NA> <NA> <NA> <NA> SimpleImputer() None OneHotEncoder()
2 0.745 0.703 0.787 0.757 44.000 745.000 33.000 6.000 0.608 gini SimpleImputer() None CustomOrdinalEncoder()
3 0.737 0.716 0.757 0.528 70.000 1,003.000 37.000 19.000 0.530 gini SimpleImputer(strategy='median') PCA('mle') CustomOrdinalEncoder()
4 0.734 0.709 0.758 0.422 68.000 1,348.000 4.000 39.000 0.859 entropy SimpleImputer(strategy='median') PCA('mle') CustomOrdinalEncoder()
5 0.727 0.694 0.759 0.940 40.000 899.000 3.000 16.000 0.696 entropy SimpleImputer(strategy='median') None CustomOrdinalEncoder()
In [19]:
results.to_formatted_dataframe(query='model == "LogisticRegression()"', include_rank=True)
Out[19]:
rank roc_auc Mean roc_auc 95CI.LO roc_auc 95CI.HI C imputer scaler pca encoder
1 0.762 0.714 0.810 0.014 SimpleImputer() StandardScaler() PCA('mle') OneHotEncoder()
2 0.759 0.712 0.806 0.000 SimpleImputer(strategy='most_frequent') MinMaxScaler() None OneHotEncoder()
3 0.757 0.711 0.803 <NA> SimpleImputer() StandardScaler() None OneHotEncoder()
4 0.729 0.645 0.814 0.000 SimpleImputer(strategy='most_frequent') StandardScaler() PCA('mle') OneHotEncoder()
5 0.723 0.705 0.741 0.000 SimpleImputer(strategy='median') StandardScaler() None CustomOrdinalEncoder()

BayesSearchCV Performance Over Time¶

In [20]:
results.plot_performance_across_trials(facet_by='model').show()
In [21]:
results.plot_performance_across_trials(query='model == "RandomForestClassifier()"').show()

Variable Performance Over Time¶

In [22]:
results.plot_parameter_values_across_trials(query='model == "RandomForestClassifier()"').show()

Scatter Matrix¶

In [23]:
# results.plot_scatter_matrix(query='model == "RandomForestClassifier()"',
#                             height=1000, width=1000).show()

Variable Performance - Numeric¶

In [24]:
results.plot_performance_numeric_params(query='model == "RandomForestClassifier()"',
                                        height=800)
In [25]:
results.plot_parallel_coordinates(query='model == "RandomForestClassifier()"').show()

Variable Performance - Non-Numeric¶

In [26]:
results.plot_performance_non_numeric_params(query='model == "RandomForestClassifier()"').show()

In [27]:
results.plot_score_vs_parameter(
    query='model == "RandomForestClassifier()"',
    parameter='max_features',
    size='max_depth',
    color='encoder',
)

In [28]:
# results.plot_parameter_vs_parameter(
#     query='model == "XGBClassifier()"',
#     parameter_x='colsample_bytree',
#     parameter_y='learning_rate',
#     size='max_depth'
# )
In [29]:
# results.plot_parameter_vs_parameter(
#     query='model == "XGBClassifier()"',
#     parameter_x='colsample_bytree',
#     parameter_y='learning_rate',
#     size='imputer'
# )

Best Model - Test Set Performance¶

In [30]:
test_predictions = best_estimator.predict(X_test)
test_predictions[0:10]
Out[30]:
array([0.32794716, 0.32893009, 0.37430263, 0.3052219 , 0.23870015,
       0.31131838, 0.24759746, 0.33864031, 0.23092148, 0.24922654])
In [31]:
evaluator = hlp.sklearn_eval.TwoClassEvaluator(
    actual_values=y_test,
    predicted_scores=test_predictions,
    score_threshold=0.37
)
In [32]:
evaluator.plot_actual_vs_predict_histogram()
In [33]:
evaluator.plot_confusion_matrix()
In [34]:
evaluator.all_metrics_df(return_style=True,
                         dummy_classifier_strategy=['prior', 'constant'],
                         round_by=3)
Out[34]:
  Score Dummy (prior) Dummy (constant) Explanation
AUC 0.781 0.500 0.500 Area under the ROC curve (true pos. rate vs false pos. rate); ranges from 0.5 (purely random classifier) to 1.0 (perfect classifier)
True Positive Rate 0.169 0.000 1.000 16.9% of positive instances were correctly identified.; i.e. 10 "Positive Class" labels were correctly identified out of 59 instances; a.k.a Sensitivity/Recall
True Negative Rate 0.957 1.000 0.000 95.7% of negative instances were correctly identified.; i.e. 135 "Negative Class" labels were correctly identified out of 141 instances
False Positive Rate 0.043 0.000 1.000 4.3% of negative instances were incorrectly identified as positive; i.e. 6 "Negative Class" labels were incorrectly identified as "Positive Class", out of 141 instances
False Negative Rate 0.831 1.000 0.000 83.1% of positive instances were incorrectly identified as negative; i.e. 49 "Positive Class" labels were incorrectly identified as "Negative Class", out of 59 instances
Positive Predictive Value 0.625 0.000 0.295 When the model claims an instance is positive, it is correct 62.5% of the time; i.e. out of the 16 times the model predicted "Positive Class", it was correct 10 times; a.k.a precision
Negative Predictive Value 0.734 0.705 0.000 When the model claims an instance is negative, it is correct 73.4% of the time; i.e. out of the 184 times the model predicted "Negative Class", it was correct 135 times
F1 Score 0.267 0.000 0.456 The F1 score can be interpreted as a weighted average of the precision and recall, where an F1 score reaches its best value at 1 and worst score at 0.
Precision/Recall AUC 0.621 0.295 0.295 Precision/Recall AUC is calculated with `average_precision` which summarizes a precision-recall curve as the weighted mean of precisions achieved at each threshold. See sci-kit learn documentation for caveats.
Accuracy 0.725 0.705 0.295 72.5% of instances were correctly identified
Error Rate 0.275 0.295 0.705 27.5% of instances were incorrectly identified
% Positive 0.295 0.295 0.295 29.5% of the data are positive; i.e. out of 200 total observations; 59 are labeled as "Positive Class"
Total Observations 200 200 200 There are 200 total observations; i.e. sample size
In [35]:
evaluator.plot_roc_auc_curve().show()
<Figure size 720x444.984 with 0 Axes>
In [36]:
evaluator.plot_precision_recall_auc_curve().show()
In [37]:
evaluator.plot_threshold_curves(score_threshold_range=(0.1, 0.7)).show()
In [38]:
evaluator.plot_precision_recall_tradeoff(score_threshold_range=(0.1, 0.6)).show()
In [39]:
evaluator.calculate_lift_gain(return_style=True)
Out[39]:
  Gain Lift
Percentile    
5 0.17 3.39
10 0.20 2.03
15 0.34 2.26
20 0.44 2.20
25 0.49 1.97
30 0.56 1.86
35 0.63 1.79
40 0.69 1.74
45 0.76 1.69
50 0.80 1.59
55 0.83 1.51
60 0.85 1.41
65 0.86 1.33
70 0.93 1.33
75 0.95 1.27
80 0.97 1.21
85 0.98 1.16
90 1.00 1.11
95 1.00 1.05
100 1.00 1.00

Production Model - Test Set Performance¶

In [40]:
test_predictions = production_model.predict(X_test)
test_predictions[0:10]
Out[40]:
array([0.32794716, 0.32893009, 0.37430263, 0.3052219 , 0.23870015,
       0.31131838, 0.24759746, 0.33864031, 0.23092148, 0.24922654])
In [41]:
evaluator = hlp.sklearn_eval.TwoClassEvaluator(
    actual_values=y_test,
    predicted_scores=test_predictions,
    score_threshold=0.37
)
In [42]:
evaluator.plot_actual_vs_predict_histogram()
In [43]:
evaluator.plot_confusion_matrix()
In [44]:
evaluator.all_metrics_df(return_style=True,
                         dummy_classifier_strategy=['prior', 'constant'],
                         round_by=3)
Out[44]:
  Score Dummy (prior) Dummy (constant) Explanation
AUC 0.781 0.500 0.500 Area under the ROC curve (true pos. rate vs false pos. rate); ranges from 0.5 (purely random classifier) to 1.0 (perfect classifier)
True Positive Rate 0.169 0.000 1.000 16.9% of positive instances were correctly identified.; i.e. 10 "Positive Class" labels were correctly identified out of 59 instances; a.k.a Sensitivity/Recall
True Negative Rate 0.957 1.000 0.000 95.7% of negative instances were correctly identified.; i.e. 135 "Negative Class" labels were correctly identified out of 141 instances
False Positive Rate 0.043 0.000 1.000 4.3% of negative instances were incorrectly identified as positive; i.e. 6 "Negative Class" labels were incorrectly identified as "Positive Class", out of 141 instances
False Negative Rate 0.831 1.000 0.000 83.1% of positive instances were incorrectly identified as negative; i.e. 49 "Positive Class" labels were incorrectly identified as "Negative Class", out of 59 instances
Positive Predictive Value 0.625 0.000 0.295 When the model claims an instance is positive, it is correct 62.5% of the time; i.e. out of the 16 times the model predicted "Positive Class", it was correct 10 times; a.k.a precision
Negative Predictive Value 0.734 0.705 0.000 When the model claims an instance is negative, it is correct 73.4% of the time; i.e. out of the 184 times the model predicted "Negative Class", it was correct 135 times
F1 Score 0.267 0.000 0.456 The F1 score can be interpreted as a weighted average of the precision and recall, where an F1 score reaches its best value at 1 and worst score at 0.
Precision/Recall AUC 0.621 0.295 0.295 Precision/Recall AUC is calculated with `average_precision` which summarizes a precision-recall curve as the weighted mean of precisions achieved at each threshold. See sci-kit learn documentation for caveats.
Accuracy 0.725 0.705 0.295 72.5% of instances were correctly identified
Error Rate 0.275 0.295 0.705 27.5% of instances were incorrectly identified
% Positive 0.295 0.295 0.295 29.5% of the data are positive; i.e. out of 200 total observations; 59 are labeled as "Positive Class"
Total Observations 200 200 200 There are 200 total observations; i.e. sample size
In [45]:
evaluator.plot_roc_auc_curve().show()
<Figure size 720x444.984 with 0 Axes>
In [46]:
evaluator.plot_precision_recall_auc_curve().show()
In [47]:
evaluator.plot_threshold_curves(score_threshold_range=(0.1, 0.7)).show()
In [48]:
evaluator.plot_precision_recall_tradeoff(score_threshold_range=(0.1, 0.6)).show()
In [49]:
evaluator.calculate_lift_gain(return_style=True)
Out[49]:
  Gain Lift
Percentile    
5 0.17 3.39
10 0.20 2.03
15 0.34 2.26
20 0.44 2.20
25 0.49 1.97
30 0.56 1.86
35 0.63 1.79
40 0.69 1.74
45 0.76 1.69
50 0.80 1.59
55 0.83 1.51
60 0.85 1.41
65 0.86 1.33
70 0.93 1.33
75 0.95 1.27
80 0.97 1.21
85 0.98 1.16
90 1.00 1.11
95 1.00 1.05
100 1.00 1.00